gpus4all
Sign in Start free
Changelog

What changed

New features, fixes, and the occasional entry about something we broke. Most changes ship to the API without a version bump because the API is additive and we keep it that way.

2026-07-02

Streaming usage counters

  • stream_options.include_usage now works on streamed chat completions; the final chunk carries real token counts.
2026-05-19

Spending caps on every deployment

  • New deployments get a default $50 monthly cap. Raising it requires confirming you actually want to spend more.
  • Cap enforcement returns 429 with a cap_reached code.
2026-04-08

Model hosting on L40S

  • 14B to 40B models can now deploy on an L40S tier at $99/mo base.
2026-03-02

Snapshot restore without an instance

  • You can now restore a snapshot into a new instance of a different size without creating the old size first.
2026-01-21

API keys can be scoped to one deployment

  • Keys now take an optional deployment scope so a key can only reach one model. Most of our customers rotate per deployment now.
2025-12-10

Instance boot time improvements

  • H100 boot time reduced from ~40 minutes to ~9 minutes. Details in the engineering blog.
2025-10-16

Per-second billing on invoices

  • Invoices now show per-second compute time rounded to the minute.
  • Fixed a rounding bug where a 30 second instance could bill a full minute twice in edge cases. Refunds issued.
2025-09-01

First model hosting release

  • OpenAI-compatible API, five starter models, per-token billing, and spending caps.
  • Docs live at the URL you are reading.